Commit
•
1351f60
1
Parent(s):
c025d41
Upload folder using huggingface_hub
Browse files
README.md
CHANGED
|
@@ -29,7 +29,7 @@ python utils/inference.py
|
|
| 29 |
|
| 30 |
# Research
|
| 31 |
|
| 32 |
-
I started training the model
|
| 33 |
|
| 34 |
Synthetic datasets have limitations for achieving great segmentation results. This is because artificial lighting, occlusion, scale or backgrounds create a gap between synthetic and real images. A "model trained solely on synthetic data generated with naïve domain randomization struggles to generalize on the real domain", see [PEOPLESANSPEOPLE: A Synthetic Data Generator for Human-Centric Computer Vision (2022)](https://arxiv.org/pdf/2112.09290).
|
| 35 |
|
|
@@ -41,5 +41,5 @@ Latest changes (05/07/2024):
|
|
| 41 |
|
| 42 |
Next steps:
|
| 43 |
|
| 44 |
-
- Expand dataset
|
| 45 |
-
- Research on multi-step segmentation by incorporating [ViTMatte](https://github.com/hustvl/ViTMatte)
|
|
|
|
| 29 |
|
| 30 |
# Research
|
| 31 |
|
| 32 |
+
I started training the model with synthetic images of the [Human Segmentation Dataset](https://huggingface.co/datasets/schirrmacher/humans) crafted with [LayerDiffuse](https://github.com/layerdiffusion/LayerDiffuse). However, I noticed that the model struggles to perform well on real images.
|
| 33 |
|
| 34 |
Synthetic datasets have limitations for achieving great segmentation results. This is because artificial lighting, occlusion, scale or backgrounds create a gap between synthetic and real images. A "model trained solely on synthetic data generated with naïve domain randomization struggles to generalize on the real domain", see [PEOPLESANSPEOPLE: A Synthetic Data Generator for Human-Centric Computer Vision (2022)](https://arxiv.org/pdf/2112.09290).
|
| 35 |
|
|
|
|
| 41 |
|
| 42 |
Next steps:
|
| 43 |
|
| 44 |
+
- Expand dataset with synthetic and real images
|
| 45 |
+
- Research on multi-step segmentation/matting by incorporating [ViTMatte](https://github.com/hustvl/ViTMatte)
|